1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * transition.c - Kernel Live Patching transition functions
4  *
5  * Copyright (C) 2015-2016 Josh Poimboeuf <jpoimboe@redhat.com>
6  */
7 
8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
9 
10 #include <linux/cpu.h>
11 #include <linux/stacktrace.h>
12 #include "core.h"
13 #include "patch.h"
14 #include "transition.h"
15 
16 #define MAX_STACK_ENTRIES  100
17 #define STACK_ERR_BUF_SIZE 128
18 
19 #define SIGNALS_TIMEOUT 15
20 
21 struct klp_patch *klp_transition_patch;
22 
23 static int klp_target_state = KLP_UNDEFINED;
24 
25 static unsigned int klp_signals_cnt;
26 
27 /*
28  * This work can be performed periodically to finish patching or unpatching any
29  * "straggler" tasks which failed to transition in the first attempt.
30  */
klp_transition_work_fn(struct work_struct * work)31 static void klp_transition_work_fn(struct work_struct *work)
32 {
33 	mutex_lock(&klp_mutex);
34 
35 	if (klp_transition_patch)
36 		klp_try_complete_transition();
37 
38 	mutex_unlock(&klp_mutex);
39 }
40 static DECLARE_DELAYED_WORK(klp_transition_work, klp_transition_work_fn);
41 
42 /*
43  * This function is just a stub to implement a hard force
44  * of synchronize_rcu(). This requires synchronizing
45  * tasks even in userspace and idle.
46  */
klp_sync(struct work_struct * work)47 static void klp_sync(struct work_struct *work)
48 {
49 }
50 
51 /*
52  * We allow to patch also functions where RCU is not watching,
53  * e.g. before user_exit(). We can not rely on the RCU infrastructure
54  * to do the synchronization. Instead hard force the sched synchronization.
55  *
56  * This approach allows to use RCU functions for manipulating func_stack
57  * safely.
58  */
klp_synchronize_transition(void)59 static void klp_synchronize_transition(void)
60 {
61 	schedule_on_each_cpu(klp_sync);
62 }
63 
64 /*
65  * The transition to the target patch state is complete.  Clean up the data
66  * structures.
67  */
klp_complete_transition(void)68 static void klp_complete_transition(void)
69 {
70 	struct klp_object *obj;
71 	struct klp_func *func;
72 	struct task_struct *g, *task;
73 	unsigned int cpu;
74 
75 	pr_debug("'%s': completing %s transition\n",
76 		 klp_transition_patch->mod->name,
77 		 klp_target_state == KLP_PATCHED ? "patching" : "unpatching");
78 
79 	if (klp_transition_patch->replace && klp_target_state == KLP_PATCHED) {
80 		klp_unpatch_replaced_patches(klp_transition_patch);
81 		klp_discard_nops(klp_transition_patch);
82 	}
83 
84 	if (klp_target_state == KLP_UNPATCHED) {
85 		/*
86 		 * All tasks have transitioned to KLP_UNPATCHED so we can now
87 		 * remove the new functions from the func_stack.
88 		 */
89 		klp_unpatch_objects(klp_transition_patch);
90 
91 		/*
92 		 * Make sure klp_ftrace_handler() can no longer see functions
93 		 * from this patch on the ops->func_stack.  Otherwise, after
94 		 * func->transition gets cleared, the handler may choose a
95 		 * removed function.
96 		 */
97 		klp_synchronize_transition();
98 	}
99 
100 	klp_for_each_object(klp_transition_patch, obj)
101 		klp_for_each_func(obj, func)
102 			func->transition = false;
103 
104 	/* Prevent klp_ftrace_handler() from seeing KLP_UNDEFINED state */
105 	if (klp_target_state == KLP_PATCHED)
106 		klp_synchronize_transition();
107 
108 	read_lock(&tasklist_lock);
109 	for_each_process_thread(g, task) {
110 		WARN_ON_ONCE(test_tsk_thread_flag(task, TIF_PATCH_PENDING));
111 		task->patch_state = KLP_UNDEFINED;
112 	}
113 	read_unlock(&tasklist_lock);
114 
115 	for_each_possible_cpu(cpu) {
116 		task = idle_task(cpu);
117 		WARN_ON_ONCE(test_tsk_thread_flag(task, TIF_PATCH_PENDING));
118 		task->patch_state = KLP_UNDEFINED;
119 	}
120 
121 	klp_for_each_object(klp_transition_patch, obj) {
122 		if (!klp_is_object_loaded(obj))
123 			continue;
124 		if (klp_target_state == KLP_PATCHED)
125 			klp_post_patch_callback(obj);
126 		else if (klp_target_state == KLP_UNPATCHED)
127 			klp_post_unpatch_callback(obj);
128 	}
129 
130 	pr_notice("'%s': %s complete\n", klp_transition_patch->mod->name,
131 		  klp_target_state == KLP_PATCHED ? "patching" : "unpatching");
132 
133 	klp_target_state = KLP_UNDEFINED;
134 	klp_transition_patch = NULL;
135 }
136 
137 /*
138  * This is called in the error path, to cancel a transition before it has
139  * started, i.e. klp_init_transition() has been called but
140  * klp_start_transition() hasn't.  If the transition *has* been started,
141  * klp_reverse_transition() should be used instead.
142  */
klp_cancel_transition(void)143 void klp_cancel_transition(void)
144 {
145 	if (WARN_ON_ONCE(klp_target_state != KLP_PATCHED))
146 		return;
147 
148 	pr_debug("'%s': canceling patching transition, going to unpatch\n",
149 		 klp_transition_patch->mod->name);
150 
151 	klp_target_state = KLP_UNPATCHED;
152 	klp_complete_transition();
153 }
154 
155 /*
156  * Switch the patched state of the task to the set of functions in the target
157  * patch state.
158  *
159  * NOTE: If task is not 'current', the caller must ensure the task is inactive.
160  * Otherwise klp_ftrace_handler() might read the wrong 'patch_state' value.
161  */
klp_update_patch_state(struct task_struct * task)162 void klp_update_patch_state(struct task_struct *task)
163 {
164 	/*
165 	 * A variant of synchronize_rcu() is used to allow patching functions
166 	 * where RCU is not watching, see klp_synchronize_transition().
167 	 */
168 	preempt_disable_notrace();
169 
170 	/*
171 	 * This test_and_clear_tsk_thread_flag() call also serves as a read
172 	 * barrier (smp_rmb) for two cases:
173 	 *
174 	 * 1) Enforce the order of the TIF_PATCH_PENDING read and the
175 	 *    klp_target_state read.  The corresponding write barrier is in
176 	 *    klp_init_transition().
177 	 *
178 	 * 2) Enforce the order of the TIF_PATCH_PENDING read and a future read
179 	 *    of func->transition, if klp_ftrace_handler() is called later on
180 	 *    the same CPU.  See __klp_disable_patch().
181 	 */
182 	if (test_and_clear_tsk_thread_flag(task, TIF_PATCH_PENDING))
183 		task->patch_state = READ_ONCE(klp_target_state);
184 
185 	preempt_enable_notrace();
186 }
187 
188 /*
189  * Determine whether the given stack trace includes any references to a
190  * to-be-patched or to-be-unpatched function.
191  */
klp_check_stack_func(struct klp_func * func,unsigned long * entries,unsigned int nr_entries)192 static int klp_check_stack_func(struct klp_func *func, unsigned long *entries,
193 				unsigned int nr_entries)
194 {
195 	unsigned long func_addr, func_size, address;
196 	struct klp_ops *ops;
197 	int i;
198 
199 	for (i = 0; i < nr_entries; i++) {
200 		address = entries[i];
201 
202 		if (klp_target_state == KLP_UNPATCHED) {
203 			 /*
204 			  * Check for the to-be-unpatched function
205 			  * (the func itself).
206 			  */
207 			func_addr = (unsigned long)func->new_func;
208 			func_size = func->new_size;
209 		} else {
210 			/*
211 			 * Check for the to-be-patched function
212 			 * (the previous func).
213 			 */
214 			ops = klp_find_ops(func->old_func);
215 
216 			if (list_is_singular(&ops->func_stack)) {
217 				/* original function */
218 				func_addr = (unsigned long)func->old_func;
219 				func_size = func->old_size;
220 			} else {
221 				/* previously patched function */
222 				struct klp_func *prev;
223 
224 				prev = list_next_entry(func, stack_node);
225 				func_addr = (unsigned long)prev->new_func;
226 				func_size = prev->new_size;
227 			}
228 		}
229 
230 		if (address >= func_addr && address < func_addr + func_size)
231 			return -EAGAIN;
232 	}
233 
234 	return 0;
235 }
236 
237 /*
238  * Determine whether it's safe to transition the task to the target patch state
239  * by looking for any to-be-patched or to-be-unpatched functions on its stack.
240  */
klp_check_stack(struct task_struct * task,const char ** oldname)241 static int klp_check_stack(struct task_struct *task, const char **oldname)
242 {
243 	static unsigned long entries[MAX_STACK_ENTRIES];
244 	struct klp_object *obj;
245 	struct klp_func *func;
246 	int ret, nr_entries;
247 
248 	ret = stack_trace_save_tsk_reliable(task, entries, ARRAY_SIZE(entries));
249 	if (ret < 0)
250 		return -EINVAL;
251 	nr_entries = ret;
252 
253 	klp_for_each_object(klp_transition_patch, obj) {
254 		if (!obj->patched)
255 			continue;
256 		klp_for_each_func(obj, func) {
257 			ret = klp_check_stack_func(func, entries, nr_entries);
258 			if (ret) {
259 				*oldname = func->old_name;
260 				return -EADDRINUSE;
261 			}
262 		}
263 	}
264 
265 	return 0;
266 }
267 
klp_check_and_switch_task(struct task_struct * task,void * arg)268 static int klp_check_and_switch_task(struct task_struct *task, void *arg)
269 {
270 	int ret;
271 
272 	if (task_curr(task) && task != current)
273 		return -EBUSY;
274 
275 	ret = klp_check_stack(task, arg);
276 	if (ret)
277 		return ret;
278 
279 	clear_tsk_thread_flag(task, TIF_PATCH_PENDING);
280 	task->patch_state = klp_target_state;
281 	return 0;
282 }
283 
284 /*
285  * Try to safely switch a task to the target patch state.  If it's currently
286  * running, or it's sleeping on a to-be-patched or to-be-unpatched function, or
287  * if the stack is unreliable, return false.
288  */
klp_try_switch_task(struct task_struct * task)289 static bool klp_try_switch_task(struct task_struct *task)
290 {
291 	const char *old_name;
292 	int ret;
293 
294 	/* check if this task has already switched over */
295 	if (task->patch_state == klp_target_state)
296 		return true;
297 
298 	/*
299 	 * For arches which don't have reliable stack traces, we have to rely
300 	 * on other methods (e.g., switching tasks at kernel exit).
301 	 */
302 	if (!klp_have_reliable_stack())
303 		return false;
304 
305 	/*
306 	 * Now try to check the stack for any to-be-patched or to-be-unpatched
307 	 * functions.  If all goes well, switch the task to the target patch
308 	 * state.
309 	 */
310 	ret = task_call_func(task, klp_check_and_switch_task, &old_name);
311 	switch (ret) {
312 	case 0:		/* success */
313 		break;
314 
315 	case -EBUSY:	/* klp_check_and_switch_task() */
316 		pr_debug("%s: %s:%d is running\n",
317 			 __func__, task->comm, task->pid);
318 		break;
319 	case -EINVAL:	/* klp_check_and_switch_task() */
320 		pr_debug("%s: %s:%d has an unreliable stack\n",
321 			 __func__, task->comm, task->pid);
322 		break;
323 	case -EADDRINUSE: /* klp_check_and_switch_task() */
324 		pr_debug("%s: %s:%d is sleeping on function %s\n",
325 			 __func__, task->comm, task->pid, old_name);
326 		break;
327 
328 	default:
329 		pr_debug("%s: Unknown error code (%d) when trying to switch %s:%d\n",
330 			 __func__, ret, task->comm, task->pid);
331 		break;
332 	}
333 
334 	return !ret;
335 }
336 
337 /*
338  * Sends a fake signal to all non-kthread tasks with TIF_PATCH_PENDING set.
339  * Kthreads with TIF_PATCH_PENDING set are woken up.
340  */
klp_send_signals(void)341 static void klp_send_signals(void)
342 {
343 	struct task_struct *g, *task;
344 
345 	if (klp_signals_cnt == SIGNALS_TIMEOUT)
346 		pr_notice("signaling remaining tasks\n");
347 
348 	read_lock(&tasklist_lock);
349 	for_each_process_thread(g, task) {
350 		if (!klp_patch_pending(task))
351 			continue;
352 
353 		/*
354 		 * There is a small race here. We could see TIF_PATCH_PENDING
355 		 * set and decide to wake up a kthread or send a fake signal.
356 		 * Meanwhile the task could migrate itself and the action
357 		 * would be meaningless. It is not serious though.
358 		 */
359 		if (task->flags & PF_KTHREAD) {
360 			/*
361 			 * Wake up a kthread which sleeps interruptedly and
362 			 * still has not been migrated.
363 			 */
364 			wake_up_state(task, TASK_INTERRUPTIBLE);
365 		} else {
366 			/*
367 			 * Send fake signal to all non-kthread tasks which are
368 			 * still not migrated.
369 			 */
370 			set_notify_signal(task);
371 		}
372 	}
373 	read_unlock(&tasklist_lock);
374 }
375 
376 /*
377  * Try to switch all remaining tasks to the target patch state by walking the
378  * stacks of sleeping tasks and looking for any to-be-patched or
379  * to-be-unpatched functions.  If such functions are found, the task can't be
380  * switched yet.
381  *
382  * If any tasks are still stuck in the initial patch state, schedule a retry.
383  */
klp_try_complete_transition(void)384 void klp_try_complete_transition(void)
385 {
386 	unsigned int cpu;
387 	struct task_struct *g, *task;
388 	struct klp_patch *patch;
389 	bool complete = true;
390 
391 	WARN_ON_ONCE(klp_target_state == KLP_UNDEFINED);
392 
393 	/*
394 	 * Try to switch the tasks to the target patch state by walking their
395 	 * stacks and looking for any to-be-patched or to-be-unpatched
396 	 * functions.  If such functions are found on a stack, or if the stack
397 	 * is deemed unreliable, the task can't be switched yet.
398 	 *
399 	 * Usually this will transition most (or all) of the tasks on a system
400 	 * unless the patch includes changes to a very common function.
401 	 */
402 	read_lock(&tasklist_lock);
403 	for_each_process_thread(g, task)
404 		if (!klp_try_switch_task(task))
405 			complete = false;
406 	read_unlock(&tasklist_lock);
407 
408 	/*
409 	 * Ditto for the idle "swapper" tasks.
410 	 */
411 	cpus_read_lock();
412 	for_each_possible_cpu(cpu) {
413 		task = idle_task(cpu);
414 		if (cpu_online(cpu)) {
415 			if (!klp_try_switch_task(task)) {
416 				complete = false;
417 				/* Make idle task go through the main loop. */
418 				wake_up_if_idle(cpu);
419 			}
420 		} else if (task->patch_state != klp_target_state) {
421 			/* offline idle tasks can be switched immediately */
422 			clear_tsk_thread_flag(task, TIF_PATCH_PENDING);
423 			task->patch_state = klp_target_state;
424 		}
425 	}
426 	cpus_read_unlock();
427 
428 	if (!complete) {
429 		if (klp_signals_cnt && !(klp_signals_cnt % SIGNALS_TIMEOUT))
430 			klp_send_signals();
431 		klp_signals_cnt++;
432 
433 		/*
434 		 * Some tasks weren't able to be switched over.  Try again
435 		 * later and/or wait for other methods like kernel exit
436 		 * switching.
437 		 */
438 		schedule_delayed_work(&klp_transition_work,
439 				      round_jiffies_relative(HZ));
440 		return;
441 	}
442 
443 	/* we're done, now cleanup the data structures */
444 	patch = klp_transition_patch;
445 	klp_complete_transition();
446 
447 	/*
448 	 * It would make more sense to free the unused patches in
449 	 * klp_complete_transition() but it is called also
450 	 * from klp_cancel_transition().
451 	 */
452 	if (!patch->enabled)
453 		klp_free_patch_async(patch);
454 	else if (patch->replace)
455 		klp_free_replaced_patches_async(patch);
456 }
457 
458 /*
459  * Start the transition to the specified target patch state so tasks can begin
460  * switching to it.
461  */
klp_start_transition(void)462 void klp_start_transition(void)
463 {
464 	struct task_struct *g, *task;
465 	unsigned int cpu;
466 
467 	WARN_ON_ONCE(klp_target_state == KLP_UNDEFINED);
468 
469 	pr_notice("'%s': starting %s transition\n",
470 		  klp_transition_patch->mod->name,
471 		  klp_target_state == KLP_PATCHED ? "patching" : "unpatching");
472 
473 	/*
474 	 * Mark all normal tasks as needing a patch state update.  They'll
475 	 * switch either in klp_try_complete_transition() or as they exit the
476 	 * kernel.
477 	 */
478 	read_lock(&tasklist_lock);
479 	for_each_process_thread(g, task)
480 		if (task->patch_state != klp_target_state)
481 			set_tsk_thread_flag(task, TIF_PATCH_PENDING);
482 	read_unlock(&tasklist_lock);
483 
484 	/*
485 	 * Mark all idle tasks as needing a patch state update.  They'll switch
486 	 * either in klp_try_complete_transition() or at the idle loop switch
487 	 * point.
488 	 */
489 	for_each_possible_cpu(cpu) {
490 		task = idle_task(cpu);
491 		if (task->patch_state != klp_target_state)
492 			set_tsk_thread_flag(task, TIF_PATCH_PENDING);
493 	}
494 
495 	klp_signals_cnt = 0;
496 }
497 
498 /*
499  * Initialize the global target patch state and all tasks to the initial patch
500  * state, and initialize all function transition states to true in preparation
501  * for patching or unpatching.
502  */
klp_init_transition(struct klp_patch * patch,int state)503 void klp_init_transition(struct klp_patch *patch, int state)
504 {
505 	struct task_struct *g, *task;
506 	unsigned int cpu;
507 	struct klp_object *obj;
508 	struct klp_func *func;
509 	int initial_state = !state;
510 
511 	WARN_ON_ONCE(klp_target_state != KLP_UNDEFINED);
512 
513 	klp_transition_patch = patch;
514 
515 	/*
516 	 * Set the global target patch state which tasks will switch to.  This
517 	 * has no effect until the TIF_PATCH_PENDING flags get set later.
518 	 */
519 	klp_target_state = state;
520 
521 	pr_debug("'%s': initializing %s transition\n", patch->mod->name,
522 		 klp_target_state == KLP_PATCHED ? "patching" : "unpatching");
523 
524 	/*
525 	 * Initialize all tasks to the initial patch state to prepare them for
526 	 * switching to the target state.
527 	 */
528 	read_lock(&tasklist_lock);
529 	for_each_process_thread(g, task) {
530 		WARN_ON_ONCE(task->patch_state != KLP_UNDEFINED);
531 		task->patch_state = initial_state;
532 	}
533 	read_unlock(&tasklist_lock);
534 
535 	/*
536 	 * Ditto for the idle "swapper" tasks.
537 	 */
538 	for_each_possible_cpu(cpu) {
539 		task = idle_task(cpu);
540 		WARN_ON_ONCE(task->patch_state != KLP_UNDEFINED);
541 		task->patch_state = initial_state;
542 	}
543 
544 	/*
545 	 * Enforce the order of the task->patch_state initializations and the
546 	 * func->transition updates to ensure that klp_ftrace_handler() doesn't
547 	 * see a func in transition with a task->patch_state of KLP_UNDEFINED.
548 	 *
549 	 * Also enforce the order of the klp_target_state write and future
550 	 * TIF_PATCH_PENDING writes to ensure klp_update_patch_state() doesn't
551 	 * set a task->patch_state to KLP_UNDEFINED.
552 	 */
553 	smp_wmb();
554 
555 	/*
556 	 * Set the func transition states so klp_ftrace_handler() will know to
557 	 * switch to the transition logic.
558 	 *
559 	 * When patching, the funcs aren't yet in the func_stack and will be
560 	 * made visible to the ftrace handler shortly by the calls to
561 	 * klp_patch_object().
562 	 *
563 	 * When unpatching, the funcs are already in the func_stack and so are
564 	 * already visible to the ftrace handler.
565 	 */
566 	klp_for_each_object(patch, obj)
567 		klp_for_each_func(obj, func)
568 			func->transition = true;
569 }
570 
571 /*
572  * This function can be called in the middle of an existing transition to
573  * reverse the direction of the target patch state.  This can be done to
574  * effectively cancel an existing enable or disable operation if there are any
575  * tasks which are stuck in the initial patch state.
576  */
klp_reverse_transition(void)577 void klp_reverse_transition(void)
578 {
579 	unsigned int cpu;
580 	struct task_struct *g, *task;
581 
582 	pr_debug("'%s': reversing transition from %s\n",
583 		 klp_transition_patch->mod->name,
584 		 klp_target_state == KLP_PATCHED ? "patching to unpatching" :
585 						   "unpatching to patching");
586 
587 	klp_transition_patch->enabled = !klp_transition_patch->enabled;
588 
589 	klp_target_state = !klp_target_state;
590 
591 	/*
592 	 * Clear all TIF_PATCH_PENDING flags to prevent races caused by
593 	 * klp_update_patch_state() running in parallel with
594 	 * klp_start_transition().
595 	 */
596 	read_lock(&tasklist_lock);
597 	for_each_process_thread(g, task)
598 		clear_tsk_thread_flag(task, TIF_PATCH_PENDING);
599 	read_unlock(&tasklist_lock);
600 
601 	for_each_possible_cpu(cpu)
602 		clear_tsk_thread_flag(idle_task(cpu), TIF_PATCH_PENDING);
603 
604 	/* Let any remaining calls to klp_update_patch_state() complete */
605 	klp_synchronize_transition();
606 
607 	klp_start_transition();
608 }
609 
610 /* Called from copy_process() during fork */
klp_copy_process(struct task_struct * child)611 void klp_copy_process(struct task_struct *child)
612 {
613 
614 	/*
615 	 * The parent process may have gone through a KLP transition since
616 	 * the thread flag was copied in setup_thread_stack earlier. Bring
617 	 * the task flag up to date with the parent here.
618 	 *
619 	 * The operation is serialized against all klp_*_transition()
620 	 * operations by the tasklist_lock. The only exception is
621 	 * klp_update_patch_state(current), but we cannot race with
622 	 * that because we are current.
623 	 */
624 	if (test_tsk_thread_flag(current, TIF_PATCH_PENDING))
625 		set_tsk_thread_flag(child, TIF_PATCH_PENDING);
626 	else
627 		clear_tsk_thread_flag(child, TIF_PATCH_PENDING);
628 
629 	child->patch_state = current->patch_state;
630 }
631 
632 /*
633  * Drop TIF_PATCH_PENDING of all tasks on admin's request. This forces an
634  * existing transition to finish.
635  *
636  * NOTE: klp_update_patch_state(task) requires the task to be inactive or
637  * 'current'. This is not the case here and the consistency model could be
638  * broken. Administrator, who is the only one to execute the
639  * klp_force_transitions(), has to be aware of this.
640  */
klp_force_transition(void)641 void klp_force_transition(void)
642 {
643 	struct klp_patch *patch;
644 	struct task_struct *g, *task;
645 	unsigned int cpu;
646 
647 	pr_warn("forcing remaining tasks to the patched state\n");
648 
649 	read_lock(&tasklist_lock);
650 	for_each_process_thread(g, task)
651 		klp_update_patch_state(task);
652 	read_unlock(&tasklist_lock);
653 
654 	for_each_possible_cpu(cpu)
655 		klp_update_patch_state(idle_task(cpu));
656 
657 	/* Set forced flag for patches being removed. */
658 	if (klp_target_state == KLP_UNPATCHED)
659 		klp_transition_patch->forced = true;
660 	else if (klp_transition_patch->replace) {
661 		klp_for_each_patch(patch) {
662 			if (patch != klp_transition_patch)
663 				patch->forced = true;
664 		}
665 	}
666 }
667